Awesome
LiDAR2Map
Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, Jianke Zhu*
This is the official implementation of LiDAR2Map: In Defense of LiDAR-Based Semantic Map Construction Using Online Camera Distillation (CVPR 2023) [Paper] [Video].
<p align="center"> <a><img src="fig/framework.png" width="90%"></a> </p>Preparation
nuScene download
Please download the whole nuScene dataset from the official website.
Environment setup
Our project is built with Pytorch >= 1.7 and revised mmdetection3d from BEVFusion.
You can install the tree-filter by:
cd ./map/model/loss/kernels/lib_tree_filter
python3 setup.py build develop
Training and Inference
To train the model from scratch, you can run:
cd ./map
bash train.sh # multi-gpu
python train_lidar2map.py # single-gpu
To inference with the obtained checkpoint, you can run:
python test.py --modelf /path/to/ckpt # single-gpu
Acknowledgements
Thanks for the pioneer work in online map learning: HDMapNet BEVFusion BEVerse
Citations
@inproceedings{wang2023lidar2map,
title={LiDAR2Map: In Defense of LiDAR-Based Semantic Map Construction Using Online Camera Distillation},
author={Wang, Song and Li, Wentong and Liu, Wenyu and Liu, Xiaolu and Zhu, Jianke},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
pages={5186--5195},
year={2023}
}